Machine learning for classificaiton, testing it, then including the testing set into a new training set?

Hi all,

I am curious. I have a classification problem, where I need machine learning to help me place signals into three categories: increasing trend, decreasing trend, and cyclic. I have split my data into 70% for training, 30% for testing.

Say I train and test the network and then achieve accurate results. Then, I receive new, unclassed signals that I would like to include for classification. Is it a common practice to include previous results to make new predictions? If so, what methods in MATLAB are appropriate?

Thank you.

[Greg Heath]

Why is there no validation data to protect against overtraining an overfit network?

Are you making sure that the number of unknown weights doesn't exceed the number of training equations?

Or are you using Bayesian Regularization via TRAINBR?

Hope this helps.

Greg

Answer 1

[Greg Heath]

The method that I used with a radar signal database that was sporadically updated over many years:

1. Extract the reflection data for each moving target into 32 dimensional vectors. Length of target reflections varied from 1 (e.g, debris) to 16 (part of a rocket booster)) and the remainder of the signal was thermal noise.

2. The data was clustered and used to create a RBF classifer similar to a combination of MATLAB's NEWRBE and NEWRB.

3. When new data became available it was compared with the closest RBF cluster. If the target was the same type as the cluster, that vector was added to the cluster and mean and radius were updated. If not, it was used to create a new cluster.

4. This was much easier than using a feedforward MLP NN which would have been required to be completely redesigned instead of just modified/updated.

Hope this helps.

Thank you for formally accepting my answer

Greg